Object Detection from Point Clouds for Driverless Vehicles

Abstract

Point clouds offer the advantage of providing direct access to 3D data. So, utilizing solely point clouds, we present an object detection technique for driverless vehicles in this study. In this method, a graph is first constructed using the k-nearest neighbour (KNN) method, and then a graph neural network module is proposed for local information extraction. Then, depending on the coordinates of points, we utilize a pillar-based projection approach to project the locally informative feature into bird’s-eye-view (BEV). After that, residual-based networks with attention mechanisms are used for the BEV features processing. The attention system is capable of assigning appropriate weights to several nearby points, hence improving detection performance. The proposed work achieves 58.49 percent and 67.90 percent mAP (mean Average Precision) for 3D Object detection and BEV detection, respectively, on the KITTI 3D object detection benchmark. We also tested the proposed method on our driverless vehicle in our campus, and we compared it to the prior method we utilized. The experimental results reveal that the proposed method outperforms the existing methods (about 2% higher than PointPillars in BEV detection) and gives more accurate information.